Skip to content

Preprint March 2026

Benchmarking Multi-Agent LLM Architectures for Home Energy Management: Real-World Tariff Validation and Cross-Model Cost-Efficiency Analysis

Author
Besnik Sulmataj
Institution
Independent
Email
besniksulmataj@gmail.com
Date
March 17, 2026
Published
Zenodo (Preprint)
  • 108 Simulations
  • 4 LLMs Benchmarked
  • 20%+ Cost Reduction (3 of 4 models)
  • $0.005 DeepSeek-V3 API cost per day

Abstract

Summary

This paper benchmarks multi-agent LLM architectures for residential home energy management using a full factorial experiment: four frontier models (Llama 4 Maverick, DeepSeek-V3, GPT-4.1, and Claude Sonnet 4.6), three US utility tariff profiles, three household archetypes, and three random seeds across 108 seven-day simulations at hourly resolution. The central finding is that LLM-driven agents reliably achieve statistically equivalent cost reductions above 20% versus an unmanaged baseline, validating their practical viability for real-world deployment. Tariff complexity proves a stronger model differentiator than household size, with performance gaps widening significantly under real-time pricing structures. When savings performance is statistically equivalent across three models, the deployment decision reduces to cost: a 17x spread in API cost across models makes this a choice about efficiency, not capability.

Methodology

The full factorial.

108 runs = 4 models × 3 tariffs × 3 household types × 3 random seeds

  • 4

    LLM Models

    Llama 4 Maverick, DeepSeek-V3, GPT-4.1, Claude Sonnet 4.6. Open-weight and proprietary frontier models.

  • 3

    US Utility Tariffs

    ComEd Hourly (real-time pricing), PG&E E-TOU-C (time-of-use), and SCE TOU-D-4 (tiered time-of-use with demand charge).

  • 3

    Household Archetypes

    Small suburban (5 kW solar, 10 kWh battery, 60 kWh EV), large suburban (10 kW solar, 20 kWh battery, 100 kWh EV), and apartment (no solar, 5 kWh battery, 40 kWh EV).

  • 3

    Random Seeds

    Each configuration is run three times with different stochastic seeds to assess variance and ensure statistical robustness of reported savings.

Results / Savings vs. Oracle Gap

Three of four clear the threshold.

Figure 1

Axis 0 to 65%

Energy Cost Savings vs. Unmanaged Baseline (%)

20% threshold

  • Llama 4 Maverick 17.4% ± 8%
  • DeepSeek-V3 37.8% ± 9%
  • GPT-4.1 44.8% ± 10%
  • Claude Sonnet 4.6 49.3% ± 11%
Energy Savings (%) versus an unmanaged baseline. Whiskers mark the reported spread on each mean. The dashed rule is the 20% threshold.

Figure 2

Axis 0 to 95%

Oracle Gap, % Above Optimal (lower is better)

  • Llama 4 Maverick 77.8% above optimal
  • DeepSeek-V3 45.9% above optimal
  • GPT-4.1 34.3% above optimal
  • Claude Sonnet 4.6 31.8% above optimal
Distance from the theoretical optimum, expressed as % above optimal. Shorter is closer to the oracle.

Table I / Household Archetypes

The three homes under test.

Table I
Parameter Small Suburban Large Suburban Apartment
Base load (avg kW) 0.5 1.2 0.3
EV capacity (kWh) 60 100 40
EV daily use (kWh) 12 15 8
Solar peak (kW) 5 10 0
Home battery (kWh) 10 20 5

Scroll the table sideways for Large Suburban and Apartment.

Figure 3 / Tariff Generalization

Tariff complexity is the differentiator.

Figure 3

Axis 0 to 100%

Energy Savings by Tariff Profile, Grouped by Model (%)

ComEd Hourly

  • Llama 4 Maverick36.6%
  • DeepSeek-V363.1%
  • GPT-4.176.2%
  • Claude Sonnet 4.684.6%

PG&E E-TOU-C

  • Llama 4 Maverick3.8%
  • DeepSeek-V326.2%
  • GPT-4.133.6%
  • Claude Sonnet 4.639.0%

SCE TOU-D-4

  • Llama 4 Maverick11.6%
  • DeepSeek-V324.0%
  • GPT-4.124.6%
  • Claude Sonnet 4.624.3%
Savings spread widest under ComEd real-time pricing and narrowest under SCE TOU-D-4. Select a model in the legend to isolate it across all three tariffs.

Cost Efficiency & Monthly Savings

Equivalent savings, a 17x spread in cost.

Figure 4

Log x, linear y

Cost-Efficiency Scatter, API Cost vs. Energy Savings

  • Llama 4 Maverick $0.001/day · 17.4% savings
  • DeepSeek-V3 $0.005/day · 37.8% savings
  • GPT-4.1 $0.039/day · 44.8% savings
  • Claude Sonnet 4.6 $0.088/day · 49.3% savings
Energy Savings (%) against Daily API Cost (USD, log scale). Vertical axis 0 to 65%.

Figure 5

Axis $0 to $100

Monthly Net Savings by Household Type (USD)

Apartment

  • Llama 4 Maverick$5/mo
  • DeepSeek-V3$30/mo
  • GPT-4.1$35/mo
  • Claude Sonnet 4.6$28/mo

Small Suburban

  • Llama 4 Maverick$15/mo
  • DeepSeek-V3$55/mo
  • GPT-4.1$60/mo
  • Claude Sonnet 4.6$60/mo

Large Suburban

  • Llama 4 Maverick$18/mo
  • DeepSeek-V3$80/mo
  • GPT-4.1$90/mo
  • Claude Sonnet 4.6$95/mo
Monthly net savings in USD per household archetype. Select a model in the legend to isolate it across all three archetypes.

Deployment Recommendations

Which model, for which home.

  • Apartment / No Solar

    39.8%

    GPT-4.1

    Best balance of savings, API cost, and latency in the top tier. Optimal for constrained setups without solar generation or large battery assets.

  • Small Suburban / Moderate Solar

    56.0%

    Claude Sonnet 4.6

    Highest savings in the 5 kW solar configuration. Reasoning depth translates to superior solar arbitrage and EV scheduling under time-of-use tariffs.

  • Large Suburban / High Solar + EV

    54.3%

    Claude Sonnet 4.6

    6.2% oracle gap, closest to theoretical optimum. Handles complex multi-asset coordination (10 kW solar, 20 kWh battery, 100 kWh EV) with the highest fidelity.

Cite This Work

The record is public.

The preprint, the method, and every figure above are published under a DOI, so the numbers can be checked against the source rather than taken on trust.

DOI 10.5281/zenodo.19074522

BibTeX
@misc{sulmataj2026benchmarking,
  author    = {Sulmataj, Besnik},
  title     = {Benchmarking Multi-Agent LLM Architectures for Home Energy Management: Real-World Tariff Validation and Cross-Model Cost-Efficiency Analysis},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19074522},
  url       = {https://doi.org/10.5281/zenodo.19074522}
}

Back to the research index

Open line

Want to argue about where this goes?